Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Numerical Optimization for Tensor Disentanglement

T0 review · 5 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Tensor disentanglement reduces to alternating between a truncated SVD and an orthogonal Procrustes step, with a degree-of-freedom formula predicting the minimal rank.

desk verdict Solid optimization paper with a useful alternating scheme and rank heuristic; RTRN Hessian has an unaddressed regularity gap and the experimental section is absent from this draft. read the letter →

arxiv 2508.19409 v1 pith:PWE6PWSA submitted 2025-08-26 math.NA cs.NAquant-ph

Julia Wei , Alec Dektor , Chungen Shen , Zaiwen Wen , Chao Yang This is my paper · ORCID
classification math.NAcs.NAquant-ph MSC 65K0565F5515A69
keywords tensornetworksdisentanglementgaugefreedomorthogonaloptimizationRiemannianalternatingminimizationmatrixranksingularvaluedecomposition
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a tensor inside a network can be re-gauged—rotated by an orthogonal matrix on a subset of its indices—so that the matrix obtained by flattening the rotated tensor is as low-rank as possible, and that finding this rotation is a solvable numerical optimization problem. It builds two solver families: Riemannian optimization on the orthogonal manifold, using explicit SVD-based formulas for the gradient and Hessian, and an alternating algorithm that switches between a truncated SVD (best low-rank approximation for fixed rotation) and an orthogonal Procrustes step (best rotation for fixed low-rank approximation). The paper also derives a degree-of-freedom count that predicts the minimal rank for a generic l×l×l×l tensor: about 0.29 l². If these results hold, tensor-network algorithms get a cheap disentangler when the target rank is known and a principled starting point when it is not.

What carries the argument

The central object is the joint cost (5.31): minimize over orthogonal Q and rank-k matrix Mk of ||A(QX) - Mk||_F. The operator A is the index reshuffle/permutation that maps the rotated tensor into an lc×rb matrix; the objective (2.1) is the tail sum of squared singular values of A(QX). The alternating minimization is the mechanism: for fixed Q, Mk is the rank-k truncated SVD; for fixed Mk, Q is the orthogonal factor of the SVD of A^{-1}(Mk)X^T (the Procrustes solution). The rank floor is supplied by counting degrees of freedom in the skew-symmetric parameterization Q=e^B, the low-rank factors Y and W, minus gauge and S-degeneracies.

What would settle it

Take a tensor X and a rotation Q such that A(QX) has two equal singular values or a zero singular value, evaluate the Hessian-vector product from (4.19)–(4.24), and compare it with a finite-difference derivative of the Riemannian gradient: the formula will fail to match if the missing distinctness/nonzero assumption matters. Separately, for random Gaussian l×l×l×l tensors with l = 3,...,20, compute the empirically minimal achievable rank by exhaustive alternating and RCG runs from many initial Q; if the minimum falls far below or above (6.42) by a margin exceeding one or two, the degree-of-fre

Watch

Extended reading notes

Core claim

The central discovery is that the disentanglement cost can be written as a function of the tail singular values of A(QX), where A reshuffles the tensor's indices, and that this cost can be minimized while respecting the orthogonal constraint. The key contribution is the alternating algorithm: given Q, the best rank-k approximation of A(QX) is a truncated SVD; given that approximation, the best orthogonal Q is the orthogonal Procrustes solution, a polar factor from an SVD. Iterating these two steps monotonically decreases the joint objective and, the paper argues, converges to a local minimum at a per-iteration cost of O((lr)^3 + (lr)^2bc). A second contribution is a parameter-free rank estim

Load-bearing premise

The second-order solver's formulas assume the singular values of the matrix being compressed are all distinct and nonzero at every step; the objective is meant to push the smallest singular values to zero, so this assumption is weakest exactly at the solution.

Editorial extensions

If this is right

  • When a target rank k is known, Algorithm 5 gives a monotone, closed-form alternating scheme whose per-iteration cost is dominated by two matrix multiplications and one SVD of size lr×lr.
  • The Riemannian gradient (Theorem 3.1) and Hessian-vector product (Eqs. 4.19–4.24) provide the ingredients for first- and second-order methods on the orthogonal manifold, and extend to entropy-like objectives by choosing the function φ.
  • The degree-of-freedom formula (6.42) gives a principled initial guess for the unknown rank k in rank-selection routines; for l×l×l×l tensors it predicts ranks just below 0.29 l².
  • The binary-search wrapper (Algorithm 6) can turn either solver into a procedure that discovers a near-minimal rank instead of requiring it as input.
  • In test tensors drawn from isometric tensor-network states, the paper reports that the alternating scheme is competitive with or more efficient than Riemannian trust-region and conjugate-gradient methods.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The rank formula is a generic counting estimate; on structured tensors (with symmetry or entanglement constraints) the true minimal rank could be lower or higher than the formula, and an independent scan over random Gaussian tensors at several l would sharpen the prediction.
  • Because the alternating algorithm's per-iteration cost is dominated by a full lr×lr SVD, replacing that SVD with a randomized or truncated factorization could scale the method to much larger bond dimensions; the paper does not test this.
  • The same framework could be applied to unitary gate decomposition and mode transformations in quantum chemistry, where the goal is also to minimize entanglement along a bipartition; the paper mentions these applications but does not demonstrate them.
  • The local-minima landscape shown for a 4×4×4×4 random tensor suggests that most local minima are close in objective value, so initialization at the identity may be practical for small tensors; a broader empirical check across random seeds would confirm this.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper formulates tensor disentangling as an optimization problem over orthogonal matrices Q acting on a tensor index, with objective (2.1) measuring the tail singular values of a matricization A(QX). It derives a Euclidean gradient formula (Theorem 3.1, Eq. (3.10)), Riemannian gradient and Hessian-vector products for the orthogonal manifold, and proposes Riemannian conjugate gradient (RCG) and Riemannian trust-region Newton (RTRN) methods. As an alternative, it introduces an alternating minimization scheme (Algorithm 5) over (Q, M_k) using truncated SVD and orthogonal Procrustes steps, a degrees-of-freedom rank estimate (Eqs. (6.40)-(6.43)), a binary-search rank-selection procedure (Algorithm 6), and a local-minima illustration. The paper claims that the alternating method is computationally competitive with Riemannian methods and that a hybrid approach is efficient, supported by experiments on random and isoTNS tensors.

Significance. If the claims are correct, the paper would provide a useful toolbox for tensor-network codes: the computational complexity estimates for gradients and Hessian-vector products are carefully laid out, the alternating method has cheap closed-form updates, and the rank estimate (6.42) gives a concrete, falsifiable prediction for the minimal disentangling rank of l x l x l x l tensors. The manuscript also credits and builds on known SVD-differentiation results. However, several load-bearing technical points need correction or qualification, and the numerical evidence in the reviewed copy is absent, so the central performance claims remain unverified.

major comments (5)
  1. [Sec. 4.2, Eq. (4.22)] The formula for Dφ′(Σ)[E] is incorrect. Since φ′(Σ) is a diagonal matrix and φ′ is applied entrywise to the singular values, its differential must be diagonal: Dφ′(Σ)[E] = diag(φ″(σ_i)(U^T A(EX)V)_{ii}). Using the full matrix U^T A(EX)V inserts spurious off-diagonal entries into the Hessian-vector product (4.21), so Algorithm 2 does not compute Hess f(Q)[E] as claimed. This is not a cosmetic issue; it affects the correctness of RTRN. The error originates from the incorrect identity DΣ = U^T A(EX)V in Theorem 3.1.
  2. [Sec. 4.2, Eqs. (4.22)-(4.23), Algorithm 2] No regularity assumption is stated. The HVP formulas divide by Σ^{-1} and by σ_j^2 - σ_i^2, so they require all singular values of A(QX) to be nonzero and distinct. The objective (2.1) is specifically designed to drive the tail singular values to zero, so the target regime is exactly where the Hessian becomes undefined or ill-conditioned. The regularization η in (4.29) is added after Algorithm 2 has already been evaluated, so it does not cure the undefined HVP. The authors must either state the regularity assumptions and analyze the degenerate case, or replace the HVP with a regularized surrogate, and support the RTRN claims with evidence.
  3. [Sec. 3, Theorem 3.1 proof, Eq. (3.13)] The proof uses the identities U^T dU = 0 and dV^T V = 0, which are false; the correct products are skew-symmetric. Consequently Eq. (3.13), DΣ(Q)[E] = U^T A(EX)V, is not correct as a matrix equation. The final gradient formula (3.10) is nevertheless correct because the extra skew-symmetric terms are annihilated under the trace when φ′(Σ)Σ is diagonal, but the proof as written is invalid. This matters beyond the theorem because the same incorrect full-matrix identity is carried into the Hessian derivation in Section 4.2.
  4. [Sec. 5, after Eq. (5.36)] The statement that monotone decrease of (5.31) 'hence converge[s] to a local minimum' is not justified. For a two-block nonconvex alternating-minimization problem, monotone decrease of the joint objective gives convergence of the objective values and, under additional regularity, stationarity of limit points; it does not by itself imply convergence of the iterates Q^(j), M_k^(j) to a local minimum. This claim should be weakened or supported by a rigorous global-convergence argument.
  5. [Sec. 7] The numerical section is not present in the reviewed copy. Section 7 breaks off after the experimental setup in Section 7.1, before any results, convergence histories, timings, or comparisons are shown. The abstract and introduction assert that the alternating method outperforms Riemannian methods and that the hybrid approach is efficient; without Section 7 these claims are unsupported. The complete numerical experiments must be included so the algorithmic claims can be checked.
minor comments (6)
  1. [Sec. 3, after Eq. (3.10)] 'Clark subdifferential' should be 'Clarke subdifferential'.
  2. [Sec. 6.1, Eq. (6.40) paragraph] The text says the degrees of freedom of Q_l and Q_r are l(l-1)/2 and r(r-l)/2 respectively; the second expression appears to be a typo for r(r-1)/2.
  3. [Sec. 5, Algorithm 5 complexity analysis] The complexity paragraph refers to 'Steps 3-5' and lists Step 3, Step 4, Step 5, but Algorithm 5 performs the SVD in Steps 4 and 5 and computes Q in Step 6. Please renumber or correct the cross-references.
  4. [Sec. 2, Eq. (2.5)] The condition σ_k < μ < σ_{k+1} should be stated relative to the singular values of A(QX) at the current or optimal Q; as written it is ambiguous.
  5. [Sec. 6.2, Algorithm 6] The text says 'kr is not a strict upper bound' while kr is used as an upper bound in the binary search. The invariant of kl, kr, and kopt should be clarified to avoid confusion.
  6. [Throughout] Algorithm cross-references are inconsistent: Algorithm 1 is sometimes called 'Algorithm 3', and the Hessian algorithm is called 'Algorithm 4.2'. Please unify the numbering.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the gradient, Hessian, alternating scheme, and rank estimate are derived in-paper; self-citations are contextual and not load-bearing.

full rationale

The paper's central derivation chain is self-contained. The objective (2.1) is defined directly from the tail singular values of A(QX); the gradient (3.10) is proved in Theorem 3.1 (and attributed to [35] as prior art, not as the sole support), and the Riemannian Hessian (4.19)-(4.24) is obtained by differentiating the SVD with explicit formulas. The alternating algorithm (5.31)-(5.36) is a standard alternating least-squares/Procrustes scheme, with the equivalence to the original problem shown in (5.31)-(5.32); no fitted parameter is later relabeled as a prediction. The rank estimate (6.40)-(6.43) is a degree-of-freedom counting necessary condition; the paper explicitly states it is a heuristic that 'does not take into account the numerical value of each element in X or the accuracy requirement' and uses it only as a starting point for choosing k in benchmarks. Thus it is not a prediction reverse-engineered from a fit. The binary-search rank routine (Algorithm 6) checks the same truncation criterion ck(Q)≤ε that defines a good rank-k disentangler in (6.44); this is algorithmic self-consistency, not circularity. The paper cites its own prior work (e.g., [20] for a Procrustes-based Rényi-2 disentangler, [21-23] for isoTNS) but these citations are contextual and do not carry the load of the paper's claims. The Hessian formulas in (4.22)-(4.23) require nonzero and distinct singular values (division by Σ and σ_j^2−σ_i^2), and this regularity condition is not stated; that is a correctness/robustness gap, not a circularity, because the derivation does not assume the conclusion. The proof of Theorem 3.1 uses the identity U^T dU=0, which is not valid for repeated singular values, but the same formula is independently established in [35] and the trace argument cancels the offending terms for this objective; again this is a proof defect, not a circular reduction. Overall, no step reduces by construction to its inputs.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No new physical entities and no data-fitted parameters. The disentangler Q is an established concept from the cited tensor-network literature. The central derivations rest on two standard results (gradient formula from [35], SVD differentiation from [42]) and two soft modeling premises (single-tensor disentangling as the network-level unit, and generic solvability of the DOF count). The most fragile premise, unstated in the paper, is the regularity required for the Hessian formulas (distinct, nonzero singular values).

free parameters (2)
  • Regularization parameter eta in RTRN-reg (Eq. 4.29) = 1e-12
    Hand-chosen in Section 7.1 to make the Newton operator positive definite. Solver hyperparameter; not fitted to data and does not carry the central claim.
  • Truncation tolerance epsilon and maxiter in Algorithm 6 = user-specified
    User-provided stopping criteria for the binary-search rank selection. Standard algorithmic knobs, not fitted to data.
assumptions (4)
  • standard math SVD differentiation formulas (4.22)-(4.23), valid when sigma_i are distinct and nonzero
    Invoked in Section 4.2 to build the Riemannian Hessian. The paper does not state the distinctness/nonsingularity condition; near a successful disentangler tail singular values go to zero and the formulas degrade.
  • standard math Gradient formula (3.10) from [35]
    Adopted as Theorem 3.1. The paper's own proof uses U^T dU = 0, which is false in general; the known result is correct.
  • domain assumption A single-tensor disentangler Q is the right building block for network-level disentangling in isoTNS
    The entire formulation (2.4), (5.31) operates on one tensor X. Its relevance to isoTNS rests on prior work [21-23] on Moses Moves and isometric tensor rings.
  • domain assumption Eq. (6.39) is solvable for generic X when the DOF count (6.40) is at least lrbc
    The degree-of-freedom counting in Section 6.1 ignores the numerical values of X and accuracy requirements; the paper itself notes the estimate can be larger than the optimal one.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Numerical Optimization for Tensor Disentanglement." pith.science (2026). https://pith.science/paper/PWE6PWSA

@misc{pith2026250819409,
  author       = {Pith},
  title        = {Pith review of: Numerical Optimization for Tensor Disentanglement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PWE6PWSA}},
  note         = {Machine review of arXiv:2508.19409}
}
read the original abstract

Tensor networks provide compact and scalable representations of high-dimensional data, enabling efficient computation in fields such as quantum physics, numerical partial differential equations (PDEs), and machine learning. This paper focuses on tensor disentangling, the task of identifying transformations that reduce bond dimensions by exploiting gauge freedom in the network. We formulate this task as an optimization problem over orthogonal matrices acting on a single tensor's indices, aiming to minimize the rank of its matricized form. We present Riemannian optimization methods and a joint optimization framework that alternates between optimizing the orthogonal transformation for a fixed low-rank approximation and optimizing the low-rank approximation for a fixed orthogonal transformation, offering a competitive alternative when the target rank is known. To seek the often unknown optimal rank, we introduce a binary search strategy integrated with the disentangling procedure. Numerical experiments on random tensors and tensors in an approximate isometric tensor network state are performed to compare different optimization methods and explore the possibility of combining different methods in a hybrid approach.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Sampling two-dimensional isometric tensor network states

    quant-ph 2026-02 conditional novelty 6.0 of 10

    Sampling from 2D isometric tensor network states is made efficient by a row-wise QR sweep and MPO-MPS contraction, with single-sample and top-K variants.

Reference graph

Works this paper leans on

54 extracted references · 24 canonical work pages · cited by 1 Pith paper

  1. [35]

    o wner , K. : \

    barticle L \"o wner , K. : \"U ber monotone Matrixfunktionen . Mathematische Zeitschrift 38 ( 1 ), 177 -- 216 ( 1934 ) 10.1007/BF01170633 barticle

    doi:10.1007/bf01170633 1934

  2. [20]

    , Pollmann , F

    barticle Zaletel , M.P. , Pollmann , F. : Isometric Tensor Network States in Two Dimensions . Physical Review Letters 124 ( 3 ), 037201 ( 2020 ) 10.1103/PhysRevLett.124.037201 barticle

    doi:10.1103/physrevlett.124.037201 2020

  3. [1]

    , Verstraete , F

    barticle Perez-Garcia , D. , Verstraete , F. , Wolf , M.M. , Cirac , J.I. : Matrix product state representations . Quantum Info. Comput. 7 ( 5 ), 401 -- 430 ( 2007 ) barticle

    work page 2007

  4. [2]

    : Tensor-train decomposition

    barticle Oseledets , I.V. : Tensor-train decomposition . SIAM Journal on Scientific Computing 33 ( 5 ), 2295 -- 2317 ( 2011 ) 10.1137/090752286 barticle

    doi:10.1137/090752286 2011

  5. [3]

    : A practical introduction to tensor networks: Matrix product states and projected entangled pair states

    barticle Or \'u s , R. : A practical introduction to tensor networks: Matrix product states and projected entangled pair states . Annals of Physics 349 , 117 -- 158 ( 2014 ) 10.1016/j.aop.2014.06.013 barticle

    doi:10.1016/j.aop.2014.06.013 2014

  6. [4]

    , P \'e rez-Garc \'i a , D

    barticle Cirac , J.I. , P \'e rez-Garc \'i a , D. , Schuch , N. , Verstraete , F. : Matrix product states and projected entangled pair states: Concepts , symmetries, theorems . Reviews of Modern Physics 93 ( 4 ), 045003 ( 2021 ) 10.1103/RevModPhys.93.045003 barticle

    doi:10.1103/revmodphys.93.045003 2021

  7. [5]

    : Class of quantum many-body states that can be efficiently simulated

    barticle Vidal , G. : Class of quantum many-body states that can be efficiently simulated . Physical Review Letters 101 ( 11 ), 110501 ( 2008 ) 10.1103/PhysRevLett.101.110501 barticle

    doi:10.1103/physrevlett.101.110501 2008

  8. [6]

    , Vidal , G

    barticle Evenbly , G. , Vidal , G. : Tensor network renormalization . Physical Review Letters 115 ( 18 ), 180405 ( 2015 ) 10.1103/PhysRevLett.115.180405 barticle

    doi:10.1103/physrevlett.115.180405 2015

Show all 54 references
  1. [7]

    , Bader , B.W

    barticle Kolda , T.G. , Bader , B.W. : Tensor decompositions and applications . SIAM Review 51 ( 3 ), 455 -- 500 ( 2009 ) 10.1137/07070111X barticle

    2009 doi

  2. [8]

    , Vandereycken , B

    barticle Uschmajew , A. , Vandereycken , B. : The geometry of algorithms using hierarchical tensors . Linear Algebra and its Applications 439 ( 1 ), 133 -- 166 ( 2013 ) 10.1016/j.laa.2013.03.016 barticle

    2013 doi

  3. [9]

    : Tensor networks for complex quantum systems

    barticle Or \'u s , R. : Tensor networks for complex quantum systems . Nature Reviews Physics 1 ( 9 ), 538 -- 550 ( 2019 ) 10.1038/s42254-019-0086-7 barticle

    2019 doi

  4. [10]

    , Schneider , R

    barticle Bachmayr , M. , Schneider , R. , Uschmajew , A. : Tensor Networks and Hierarchical Tensors for the Solution of High-Dimensional Partial Differential Equations . Foundations of Computational Mathematics 16 ( 6 ), 1423 -- 1472 ( 2016 ) 10.1007/s10208-016-9317-9 barticle

    2016 doi

  5. [11]

    , Rodgers , A

    barticle Dektor , A. , Rodgers , A. , Venturi , D. : Rank- Adaptive Tensor Methods for High-Dimensional Nonlinear PDEs . Journal of Scientific Computing 88 ( 2 ), 36 ( 2021 ) 10.1007/s10915-021-01539-3 barticle

    2021 doi

  6. [12]

    : Tensor networks for big data analytics and large-scale optimization problems

    botherref Cichocki , A. : Tensor networks for big data analytics and large-scale optimization problems. arXiv preprint arXiv:1407.3124 (2014) botherref

    2014 arXiv

  7. [13]

    : Learning relevant features of data with multi-scale tensor networks

    barticle Stoudenmire , E.M. : Learning relevant features of data with multi-scale tensor networks . Quantum Science and Technology 3 ( 3 ), 034003 ( 2018 ) 10.1088/2058-9565/aaba1a barticle

    2018 doi

  8. [14]

    , Schotth \"o fer , S

    barticle Zangrando , E. , Schotth \"o fer , S. , Ceruti , G. , Kusch , J. , Tudisco , F. : Geometry-aware training of factorized layers in tensor tucker format . Advances in Neural Information Processing Systems 37 , 129743 -- 129773 ( 2024 ) barticle

    2024

  9. [15]

    : The density-matrix renormalization group in the age of matrix product states

    barticle Schollw \"o ck , U. : The density-matrix renormalization group in the age of matrix product states . Annals of Physics 326 ( 1 ), 96 -- 192 ( 2011 ) 10.1016/j.aop.2010.09.012 barticle

    2011 doi

  10. [16]

    , Vidal , G

    barticle Evenbly , G. , Vidal , G. : Algorithms for entanglement renormalization . Physical Review B 79 ( 14 ), 144108 ( 2009 ) 10.1103/PhysRevB.79.144108 barticle

    2009 doi

  11. [17]

    , Vidal , G

    barticle Evenbly , G. , Vidal , G. : Class of highly entangled many-body states that can be efficiently simulated . Physical Review Letters 112 ( 24 ), 240502 ( 2014 ) 10.1103/PhysRevLett.112.240502 barticle

    2014 doi

  12. [18]

    , Bardarson , J.H

    barticle Karrasch , C. , Bardarson , J.H. , Moore , J.E. : Reducing the numerical effort of finite-temperature density matrix renormalization group calculations . New Journal of Physics 15 ( 8 ), 083031 ( 2013 ) 10.1088/1367-2630/15/8/083031 barticle

    2013 doi

  13. [19]

    , Leviatan , E

    barticle Hauschild , J. , Leviatan , E. , Bardarson , J.H. , Altman , E. , Zaletel , M.P. , Pollmann , F. : Finding purifications with minimal entanglement . Physical Review B 98 ( 23 ), 235163 ( 2018 ) 10.1103/PhysRevB.98.235163 barticle

    2018 doi

  14. [21]

    , Zaletel , M.P

    barticle Lin , S.-H. , Zaletel , M.P. , Pollmann , F. : Efficient simulation of dynamics in two-dimensional quantum spin systems with isometric tensor networks . Physical Review B 106 ( 24 ), 245102 ( 2022 ) 10.1103/PhysRevB.106.245102 barticle

    2022 doi

  15. [22]

    , Kawano , M

    botherref Sappler , B. , Kawano , M. , Zaletel , M.P. , Pollmann , F. : Diagonal isometric form for tensor product states in two dimensions (arXiv:2507.08080) (2025) 10.48550/arXiv.2507.08080 botherref

    2025 doi arXiv

  16. [23]

    : Fast Tensor Disentangling Algorithm

    barticle Slagle , K. : Fast Tensor Disentangling Algorithm . SciPost Physics 11 ( 3 ), 056 ( 2021 ) 10.21468/SciPostPhys.11.3.056 barticle

    2021 doi

  17. [24]

    , Jahromi , S.S

    botherref Aizpurua , B. , Jahromi , S.S. , Singh , S. , Orus , R. : Quantum large language models via tensor network disentanglers. arXiv preprint arXiv:2410.17397 (2024) botherref

    2024 arXiv

  18. [25]

    , Veis , L

    barticle Krumnow , C. , Veis , L. , Legeza , \"O . , Eisert , J. : Fermionic orbital optimization in tensor network states . Physical Review Letters 117 ( 21 ), 210402 ( 2016 ) 10.1103/PhysRevLett.117.210402 barticle

    2016 doi

  19. [26]

    , Venturi , D

    barticle Dektor , A. , Venturi , D. : Tensor rank reduction via coordinate flows . Journal of Computational Physics 491 , 112378 ( 2023 ) 10.1016/j.jcp.2023.112378 barticle

    2023

  20. [27]

    , Venturi , D

    barticle Dektor , A. , Venturi , D. : Coordinate- Adaptive Integration of PDEs on Tensor Manifolds . Communications on Applied Mathematics and Computation 7 ( 4 ), 1562 -- 1579 ( 2025 ) 10.1007/s42967-023-00357-8 barticle

    2025 doi

  21. [28]

    , Van Damme , M

    barticle Hauru , M. , Van Damme , M. , Haegeman , J. : Riemannian optimization of isometric tensor networks . SciPost Physics 10 ( 2 ), 040 ( 2021 ) 10.21468/SciPostPhys.10.2.040 barticle

    2021 doi

  22. [29]

    , Ryzhov , A

    barticle Luchnikov , I. , Ryzhov , A. , Filippov , S. , Ouerdane , H. : QGOpt : Riemannian optimization for quantum technologies . SciPost Physics 10 ( 3 ), 079 ( 2021 ) 10.21468/SciPostPhys.10.3.079 barticle

    2021 doi

  23. [30]

    , Chuang , I.L

    botherref Nielsen , M.A. , Chuang , I.L. : Quantum Computation and Quantum Information : 10th Anniversary Edition . Cambridge University Press (2010). 10.1017/CBO9780511976667 botherref

    2010 doi

  24. [31]

    , Cirac , J.I

    barticle Verstraete , F. , Cirac , J.I. : Matrix product states represent ground states faithfully . Physical Review B 73 ( 9 ), 094423 ( 2006 ) 10.1103/PhysRevB.73.094423 barticle

    2006 doi

  25. [32]

    , Fazel , M

    barticle Recht , B. , Fazel , M. , Parrilo , P.A. : Guaranteed minimum-rank solutions of linear matrix equations via nuclear norm minimization . SIAM Review 52 ( 3 ), 471 -- 501 ( 2010 ) 10.1137/070697835 barticle

    2010 doi

  26. [33]

    , Tang , J

    barticle Lu , C. , Tang , J. , Yan , S. , Lin , Z. : Nonconvex nonsmooth low rank minimization via iteratively reweighted nuclear norm . IEEE Transactions on Image Processing 25 ( 2 ), 829 -- 839 ( 2015 ) 10.1109/TIP.2015.2511584 barticle

    2015

  27. [34]

    , Sendov , H.S

    barticle Lewis , A.S. , Sendov , H.S. : Nonsmooth Analysis of Singular Values . Part II : Applications . Set-Valued Analysis 13 ( 3 ), 243 -- 264 ( 2005 ) 10.1007/s11228-004-7198-6 . Accessed 2025-08-21 barticle

    2005 doi

  28. [36]

    , Sun , D

    barticle Ding , C. , Sun , D. , Sun , J. , Toh , K.-C. : Spectral operators of matrices . Mathematical Programming 168 ( 1 ), 509 -- 531 ( 2018 ) 10.1007/s10107-017-1162-3 barticle

    2018 doi

  29. [37]

    , Sendov , H.S

    barticle Lewis , A.S. , Sendov , H.S. : Nonsmooth Analysis of Singular Values . Part I : Theory . Set-Valued Analysis 13 ( 3 ), 213 -- 241 ( 2005 ) 10.1007/s11228-004-7197-7 barticle

    2005 doi

  30. [38]

    , Mahony , R

    bbook Absil , P.-A. , Mahony , R. , Sepulchre , R. : Optimization Algorithms on Matrix Manifolds . Princeton University Press , Princeton, New Jersey ( 2007 ) bbook

    2007

  31. [39]

    : An Introduction to Optimization on Smooth Manifolds

    bbook Boumal , N. : An Introduction to Optimization on Smooth Manifolds . Cambridge University Press , Cambridge, UK ( 2023 ) bbook

    2023

  32. [40]

    : Development of a diagonal isometric tensor network algorithm

    botherref Sappler , B. : Development of a diagonal isometric tensor network algorithm. Master's thesis, Technische Universität München (2024) botherref

    2024

  33. [41]

    : Differentiating the Singular Value Decomposition

    botherref Townsend , J. : Differentiating the Singular Value Decomposition. https://j-towns.github.io/papers/svd-derivative.pdf botherref

  34. [42]

    : Riemannian conjugate gradient methods: General framework and specific algorithms with convergence analyses

    barticle Sato , H. : Riemannian conjugate gradient methods: General framework and specific algorithms with convergence analyses . SIAM Journal on Optimization 32 ( 4 ), 2690 -- 2717 ( 2022 ) 10.1137/21M1464178 barticle

    2022 doi

  35. [43]

    , Reeves , C.M

    barticle Fletcher , R. , Reeves , C.M. : Function minimization by conjugate gradients . Comput. J. 7 , 149 -- 154 ( 1964 ) barticle

    1964

  36. [44]

    , Ribi \`e re , G

    barticle Polak , E. , Ribi \`e re , G. : Note sur la convergence de m \'e thodes de directions conjugu \'e es . ESAIM: Mathematical Modelling and Numerical Analysis - Mod \'e lisation Math \'e matique et Analyse Num \'e rique 3 ( R1 ), 35 -- 43 ( 1969 ) barticle

    1969

  37. [45]

    : The conjugate gradient method in extremal problems

    barticle Polyak , B.T. : The conjugate gradient method in extremal problems . USSR Computational Mathematics and Mathematical Physics 9 ( 4 ), 94 -- 112 ( 1969 ) 10.1016/0041-5553(69)90035-4 barticle

    1969 doi

  38. [46]

    , Stiefel , E

    barticle Hestenes , M.R. , Stiefel , E. : Methods of conjugate gradients for solving linear systems . Journal of Research of the Bureau of National Standards 49 ( 6 ), 409 -- 436 ( 1952 ) barticle

    1952

  39. [47]

    , Yuan , Y

    barticle Dai , Y.H. , Yuan , Y. : A nonlinear conjugate gradient method with a strong global convergence property . SIAM Journal on Optimization 10 ( 1 ), 177 -- 182 ( 1999 ) 10.1137/S1052623497318992 barticle

    1999 doi

  40. [48]

    , Storey , C

    barticle Liu , Y. , Storey , C. : Efficient generalized conjugate gradient algorithms, part 1: Theory . Journal of Optimization Theory and Applications 69 ( 1 ), 129 -- 137 ( 1991 ) 10.1007/BF00940464 barticle

    1991 doi

  41. [49]

    , Gould , N.I.M

    bbook Conn , A.R. , Gould , N.I.M. , Toint , P.L. : Trust Region Methods . Society for Industrial and Applied Mathematics , Philadelphia, PA ( 2000 ). 10.1137/1.9780898719857 bbook

    doi

  42. [50]

    : The conjugate gradient method and trust regions in large scale optimization

    barticle Steihaug , T. : The conjugate gradient method and trust regions in large scale optimization . SIAM Journal on Numerical Analysis 20 ( 3 ), 626 -- 637 ( 1983 ) 10.1137/0720042 barticle

    1983 doi

  43. [51]

    , Cattell , R.B

    barticle Hurley , J.R. , Cattell , R.B. : The procrustes program: Producing direct rotation to test a hypothesized factor structure . Behavioral Science 7 ( 2 ), 258 -- 262 ( 1962 ) 10.1002/bs.3830070216 barticle

    1962 doi

  44. [52]

    , Karaman , S

    barticle Mickelin , O. , Karaman , S. : On algorithms for and computing with the tensor ring decomposition . Numerical Linear Algebra with Applications 27 ( 3 ), 2289 ( 2020 ) 10.1002/nla.2289 barticle

    2020 doi

  45. [53]

    , Koep , N

    barticle Townsend , J. , Koep , N. , Weichwald , S. : Pymanopt: A python toolbox for optimization on manifolds using automatic differentiation . Journal of Machine Learning Research 17 ( 137 ), 1 -- 5 ( 2016 ) barticle

    2016

  46. [54]

    write newline

    " write newline "" before.all 'output.state := FUNCTION string.to.integer 't := t text.length 'k := #1 'char.num := t char.num #1 substring 's := s is.num s "." = or char.num k = not and char.num #1 + 'char.num := while char.num #1 - 'char.num := t #1 char.num substring FUNCTI...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.